fde-framework 0.1.4__tar.gz → 0.1.6__tar.gz

{fde_framework-0.1.4 → fde_framework-0.1.6}/CHANGELOG.md

@@ -5,6 +5,65 @@ the project is pre-release, so everything sits under 0.1.0 until the first tag.
 
 ## [Unreleased]
 
+## [0.1.6] — 2026-09-10
+
+Found by the first full demonstration engagement — 626 real scanned
+receipts (SROIE) run through the complete lifecycle on the published
+wheel:
+
+- `fde build <bare-name>` resolved the engagement but read the sample
+  pairs from the raw argument's path, silently emitting a project with an
+  empty golden set while sixty verified pairs sat on the engagement. The
+  harness's empty-exam refusal caught it at runtime; now the path is
+  resolved once, and the build receipt prints its own exam counts
+  ("evals: 42 golden, …") with a pointed warning when pairs exist but
+  none reached the golden set.
+- The local-engine choice is a concurrency decision, not only a hardware
+  one: `fde scan <engagement>` now reads the recorded arrival rate, and
+  past ~1000 requests/day (where the gap between requests drops under a
+  generation's length) it names the sequential-queue ceiling and the
+  continuous-batching answer. The vllm and ollama stack entries carry the
+  doctrine: paged KV admits more concurrent sequences, the scheduler
+  retires requests mid-batch, and the crossover is measured, never guessed.
+
+## [0.1.5] — 2026-09-10
+
+Three independent fresh-eyes auditors ran against 0.1.4 the day it shipped
+— an adversarial code review, a stranger following only the public docs,
+and a claims audit. Everything they caught, in one release:
+
+- The retrieval eval now grades every realization shape the registry can
+  emit: a wired module-level instance wins over a fresh construction (a
+  working deployment no longer scores 0% with the index blamed), class
+  `run(query, top_k=…)` shapes are called correctly, and path-contract
+  graph-retrieval is excluded from the recall gate it could never pass.
+  Recall math hardened: duplicate relevant ids no longer inflate, wrong-
+  shaped results are errors rather than silent zeros diluting a green CI.
+- `graph-retrieval.qdrant`'s expansion loop yielded the wrong variable
+  since v0.1.3 — the graph walk was dead code and the approach silently
+  degraded to vector search. Fixed and pinned by executing the template.
+  The graph-expanded variant survives adjacency entries for deleted
+  documents (the continuous-churn case it exists for) and all three
+  variants cap expansion breadth, not just depth.
+- The emitted `ops/slo.md` now shows the captured baseline as the numbers
+  to beat — it said "Not captured" over a recorded baseline, contradicting
+  its own acceptance protocol. `evals/acceptance.md` stops citing a named
+  eval owner when the client_readiness gate was waived, and stops citing
+  "those 0" cases over an empty golden set.
+- The architecture fingerprint now includes the topology: two engagements
+  building byte-different projects (different boundaries) no longer share
+  a fingerprint.
+- The interview no longer re-asks a multi-valued question it just
+  accepted; the offline_evaluability remedy names its clearing command;
+  a rejection whose applies-condition references an unanswered dimension
+  now names the question that could reverse it.
+- Docs truth pass: the quickstart names all three gate steps before the
+  passing build; the emitted-project tree shows `ops/diagnosis.md` and
+  `evals/retrieval.py`; kb commands documented registry-free as they ship;
+  README nav links absolute so they resolve on PyPI; ARCHITECTURE.md
+  diagram says seven gates; CITATION.cff current; CI checks out full
+  history so the sanitisation scan covers what the README says it covers.
+
 ## [0.1.4] — 2026-09-09
 
 The measurement release: claims the framework already made, turned into

{fde_framework-0.1.4 → fde_framework-0.1.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: fde-framework
-Version: 0.1.4
+Version: 0.1.6
 Summary: A framework for Forward Deployed Engineers: from problem statement to a deployable AI project, every decision traced to a fact.
 Project-URL: Homepage, https://github.com/atulkapoor/fde-framework
 Project-URL: Repository, https://github.com/atulkapoor/fde-framework
@@ -53,9 +53,9 @@ Description-Content-Type: text/markdown
 <p>
   <a href="#install">Install</a> ·
   <a href="#try-it">Quickstart</a> ·
-  <a href="ARCHITECTURE.md">Architecture</a> ·
-  <a href="examples/invoice-extraction/">Worked example</a> ·
-  <a href="CONTRIBUTING.md">Contributing</a> ·
+  <a href="https://github.com/atulkapoor/fde-framework/blob/main/ARCHITECTURE.md">Architecture</a> ·
+  <a href="https://github.com/atulkapoor/fde-framework/tree/main/examples/invoice-extraction">Worked example</a> ·
+  <a href="https://github.com/atulkapoor/fde-framework/blob/main/CONTRIBUTING.md">Contributing</a> ·
   <a href="https://pypi.org/project/fde-framework/">PyPI</a> ·
   <a href="https://atulkapoor.github.io/fde-framework/">Website</a>
 </p>
@@ -75,9 +75,9 @@ including on-premise, inside a customer VPC, or fully air-gapped.
 pip install fde-framework
 fde start acme --statement "Extract fields from supplier invoices."
 fde ask acme --role admin        # role-scoped discovery interview
-fde architect acme               # the design, with cited rationale
+fde architect acme               # topology + chosen approaches (rationale lands in ARCHITECTURE.md)
 fde build acme --out project     # refuses: seven gates guard the build
-# ...verify data access, capture the baseline (each gate prints its remedy), then:
+# ...verify data access, capture the baseline, name the eval owner (each gate prints its remedy), then:
 fde build acme --out project     # code + evals + deploy assets + runbook
 ```
 
@@ -131,9 +131,9 @@ pairs, role-scoped interview, hardware scan) → fact log with provenance →
 permutation space → seven gates → decide → architect → build (code, evals,
 deploy and ops assets, `RISKS.md`, `COMPLIANCE.md`) → retro and case
 capture. Overrides are honoured on the next run, trigger observations feed
-calibration, and a reviewed case can enter the corpus. over 800 tests; three
-adversarial review rounds (108 findings, each fixed and pinned as a
-regression test); CI gates on the suite, lint, and a sanitisation scan of
+calibration, and a reviewed case can enter the corpus. over 870 tests; four
+fresh-eyes review rounds (142 findings, each resolved and the fixes pinned
+as regression tests); CI gates on the suite, lint, and a sanitisation scan of
 the tree *and its history*; the evidence corpus is anchored to publicly
 documented production deployments; every decision is reproducible from its
 inputs.
@@ -265,8 +265,8 @@ fde scan engagements/acme                  # what this hardware runs
 fde cost --requests-per-day 500000 --model-b 70   # dated fleet sizing
 
 fde kb validate   # parse and cross-link the registry
-fde kb gaps     --root framework   # what the corpus is missing
-fde kb sweep    --root framework   # profiles no approach can serve
+fde kb gaps                        # what the corpus is missing
+fde kb sweep                       # profiles no approach can serve
 .venv/bin/pytest -q
 ```
 
@@ -290,10 +290,11 @@ project/
 │   └── llm.py            #   the one model touchpoint — when a decision needs a model (boundary-gated)
 ├── evals/                # golden / edge / adversarial sets from the client's own pairs
 │   ├── harness.py        #   fails CI until implemented; judge-based when the evaluation decided judged
+│   ├── retrieval.py      #   recall@10/50 of the retrieval layer alone — when retrieval answers ranked queries
 │   ├── acceptance.md     #   blind UAT protocol for the client's own judges
 │   └── load.py           #   p95 against the stated budget (when one was stated)
 ├── deploy/               # the substrate that was earned + TEARDOWN.md for all of it
-├── ops/                  # runbook keyed to the failure taxonomy, SLOs, rollback
+├── ops/                  # runbook keyed to the failure taxonomy, diagnosis walk, SLOs, rollback
 ├── ARCHITECTURE.md       # scope read-out, decisions, tools & alternatives, agent posture
 ├── RISKS.md              # every waived gate and overridden recommendation
 └── COMPLIANCE.md         # jurisdiction obligations, when a locale pack was applied
@@ -327,7 +328,7 @@ fde security-review acme --note "client infosec walked the data paths"
 fde ask acme --role eval_owner       # or: fde waive acme client_readiness --reason "..."
 
 fde build acme --out project         # refuses until the gates truly pass
-python project/evals/harness.py      # red until implemented -- that's the exam
+python project/evals/harness.py      # red: empty until pairs are seeded, then red until implemented -- that's the exam
 fde implement project                # drive a coding agent until it's green
 ```
 
@@ -389,8 +390,9 @@ the same receipts the emitted `ARCHITECTURE.md` prints.
 
 ## Troubleshooting
 
-**`no registry here`** — the default `--registry framework` is relative; run
-from the repository root or pass the path.
+**`no registry here`** — you passed `--registry`/`--root` at a directory that
+holds no registry. Drop the flag (the corpus ships inside the package) or
+point it at the `framework/` of a source checkout.
 
 **`build` refuses with gates listed** — that is the point. `fde status`
 names each gate and its remedy; soft gates take `fde waive <gate> --reason`,
@@ -546,7 +548,7 @@ The registry is the shared asset; engagements are private working state.
 ```bash
 git clone https://github.com/atulkapoor/fde-framework.git && cd fde-framework
 python3 -m venv .venv && .venv/bin/pip install -e ".[dev,documents]"
-.venv/bin/pytest -q          # ~850 tests, < 30s
+.venv/bin/pytest -q          # ~880 tests, < 60s
 .venv/bin/ruff check src tests
 ```
 

{fde_framework-0.1.4 → fde_framework-0.1.6}/README.md

@@ -15,9 +15,9 @@
 <p>
   <a href="#install">Install</a> ·
   <a href="#try-it">Quickstart</a> ·
-  <a href="ARCHITECTURE.md">Architecture</a> ·
-  <a href="examples/invoice-extraction/">Worked example</a> ·
-  <a href="CONTRIBUTING.md">Contributing</a> ·
+  <a href="https://github.com/atulkapoor/fde-framework/blob/main/ARCHITECTURE.md">Architecture</a> ·
+  <a href="https://github.com/atulkapoor/fde-framework/tree/main/examples/invoice-extraction">Worked example</a> ·
+  <a href="https://github.com/atulkapoor/fde-framework/blob/main/CONTRIBUTING.md">Contributing</a> ·
   <a href="https://pypi.org/project/fde-framework/">PyPI</a> ·
   <a href="https://atulkapoor.github.io/fde-framework/">Website</a>
 </p>
@@ -37,9 +37,9 @@ including on-premise, inside a customer VPC, or fully air-gapped.
 pip install fde-framework
 fde start acme --statement "Extract fields from supplier invoices."
 fde ask acme --role admin        # role-scoped discovery interview
-fde architect acme               # the design, with cited rationale
+fde architect acme               # topology + chosen approaches (rationale lands in ARCHITECTURE.md)
 fde build acme --out project     # refuses: seven gates guard the build
-# ...verify data access, capture the baseline (each gate prints its remedy), then:
+# ...verify data access, capture the baseline, name the eval owner (each gate prints its remedy), then:
 fde build acme --out project     # code + evals + deploy assets + runbook
 ```
 
@@ -93,9 +93,9 @@ pairs, role-scoped interview, hardware scan) → fact log with provenance →
 permutation space → seven gates → decide → architect → build (code, evals,
 deploy and ops assets, `RISKS.md`, `COMPLIANCE.md`) → retro and case
 capture. Overrides are honoured on the next run, trigger observations feed
-calibration, and a reviewed case can enter the corpus. over 800 tests; three
-adversarial review rounds (108 findings, each fixed and pinned as a
-regression test); CI gates on the suite, lint, and a sanitisation scan of
+calibration, and a reviewed case can enter the corpus. over 870 tests; four
+fresh-eyes review rounds (142 findings, each resolved and the fixes pinned
+as regression tests); CI gates on the suite, lint, and a sanitisation scan of
 the tree *and its history*; the evidence corpus is anchored to publicly
 documented production deployments; every decision is reproducible from its
 inputs.
@@ -227,8 +227,8 @@ fde scan engagements/acme                  # what this hardware runs
 fde cost --requests-per-day 500000 --model-b 70   # dated fleet sizing
 
 fde kb validate   # parse and cross-link the registry
-fde kb gaps     --root framework   # what the corpus is missing
-fde kb sweep    --root framework   # profiles no approach can serve
+fde kb gaps                        # what the corpus is missing
+fde kb sweep                       # profiles no approach can serve
 .venv/bin/pytest -q
 ```
 
@@ -252,10 +252,11 @@ project/
 │   └── llm.py            #   the one model touchpoint — when a decision needs a model (boundary-gated)
 ├── evals/                # golden / edge / adversarial sets from the client's own pairs
 │   ├── harness.py        #   fails CI until implemented; judge-based when the evaluation decided judged
+│   ├── retrieval.py      #   recall@10/50 of the retrieval layer alone — when retrieval answers ranked queries
 │   ├── acceptance.md     #   blind UAT protocol for the client's own judges
 │   └── load.py           #   p95 against the stated budget (when one was stated)
 ├── deploy/               # the substrate that was earned + TEARDOWN.md for all of it
-├── ops/                  # runbook keyed to the failure taxonomy, SLOs, rollback
+├── ops/                  # runbook keyed to the failure taxonomy, diagnosis walk, SLOs, rollback
 ├── ARCHITECTURE.md       # scope read-out, decisions, tools & alternatives, agent posture
 ├── RISKS.md              # every waived gate and overridden recommendation
 └── COMPLIANCE.md         # jurisdiction obligations, when a locale pack was applied
@@ -289,7 +290,7 @@ fde security-review acme --note "client infosec walked the data paths"
 fde ask acme --role eval_owner       # or: fde waive acme client_readiness --reason "..."
 
 fde build acme --out project         # refuses until the gates truly pass
-python project/evals/harness.py      # red until implemented -- that's the exam
+python project/evals/harness.py      # red: empty until pairs are seeded, then red until implemented -- that's the exam
 fde implement project                # drive a coding agent until it's green
 ```
 
@@ -351,8 +352,9 @@ the same receipts the emitted `ARCHITECTURE.md` prints.
 
 ## Troubleshooting
 
-**`no registry here`** — the default `--registry framework` is relative; run
-from the repository root or pass the path.
+**`no registry here`** — you passed `--registry`/`--root` at a directory that
+holds no registry. Drop the flag (the corpus ships inside the package) or
+point it at the `framework/` of a source checkout.
 
 **`build` refuses with gates listed** — that is the point. `fde status`
 names each gate and its remedy; soft gates take `fde waive <gate> --reason`,
@@ -508,7 +510,7 @@ The registry is the shared asset; engagements are private working state.
 ```bash
 git clone https://github.com/atulkapoor/fde-framework.git && cd fde-framework
 python3 -m venv .venv && .venv/bin/pip install -e ".[dev,documents]"
-.venv/bin/pytest -q          # ~850 tests, < 30s
+.venv/bin/pytest -q          # ~880 tests, < 60s
 .venv/bin/ruff check src tests
 ```
 

{fde_framework-0.1.4 → fde_framework-0.1.6}/examples/invoice-extraction/README.md

@@ -45,7 +45,7 @@ fde security-review engagements/acme --note "client infosec reviewed data paths
 fde waive engagements/acme client_readiness --reason "eval owner named, starts Monday"
 
 fde architect engagements/acme
-# topology on-prem   [82d56c303de52199]
+# topology on-prem   [de84a4cea7cc42b9]
 #   deployment       systemd-unit via plain-python
 #   evaluation       field-match via plain-python
 #   governance       boundary-and-audit via plain-python
@@ -57,6 +57,7 @@ fde architect engagements/acme
 
 fde build engagements/acme --out project
 # wrote project
+# next: fde implement project --holdout engagements/acme/artifacts/holdout.jsonl
 ```
 
 The emitted `project/` holds `app/` (pipeline in topological order, boundary
@@ -64,7 +65,7 @@ check imported at startup because data cannot leave), `evals/` (golden set
 from the three pairs, a harness that fails CI until the pipeline is
 implemented, and `acceptance.md` — a blind-judging protocol for the client's
 own people), `deploy/` (a systemd unit — rung zero, because nothing in the
-profile earned a container), `ops/` (runbook, SLOs, rollback),
+profile earned a container), `ops/` (runbook, diagnosis walk, SLOs, rollback),
 `ARCHITECTURE.md` with the scope read-out, the tools table with
 in-topology alternatives, the agent posture, and every rejected
 alternative, and `RISKS.md` recording the one waived gate and its reason.
@@ -82,7 +83,7 @@ fde retro engagements/acme --outcome "delivered" --days 18
 Notice what the decisions did *not* do: no vector database for a lookup
 workload, no Kubernetes for a single service, no LLM in the extraction path
 while the deterministic mapper's coverage is unmeasured. The fingerprint
-`82d56c303de52199` is stable — rebuild from the same facts and the diff is
+`de84a4cea7cc42b9` is stable — rebuild from the same facts and the diff is
 empty. (The corpus evolves, and a corpus change that moves a decision moves
 this fingerprint with it; the transcript above is re-run against the corpus
 it ships with.)

{fde_framework-0.1.4 → fde_framework-0.1.6}/framework/stacks/ollama.md

@@ -17,3 +17,9 @@ Not a fleet server: no continuous batching, no paged attention. The moment
 there are ten concurrent users, this is the wrong stack and vLLM (or SGLang,
 where requests share long prefixes) is the right one -- which is a swap of
 realization, not of architecture.
+
+The ceiling is concurrency: generations serve from a sequential queue, so
+overlapping requests stack their latencies. Right for a seat or a team;
+past roughly a thousand requests a day the queue is the system -- move to
+an engine with continuous batching before promising that rate, and
+`fde scan <engagement>` says so from the recorded arrival rate.

fde_framework-0.1.6/framework/stacks/vllm.md

@@ -0,0 +1,21 @@
+---
+id: vllm
+name: vLLM
+licence: Apache-2.0
+topologies: [customer-vpc, on-prem, air-gapped, hybrid]
+last_verified: 2026-09-10
+reversibility: moderate
+---
+Serving open weights. Continuous batching and prefix caching by default.
+
+What the defaults buy: the KV cache is paged rather than allocated as one
+contiguous block per request, so memory fragments less and admits more
+concurrent sequences; and the scheduler adds and retires requests mid-batch
+instead of waiting for a batch to drain. Together they are why this engine
+holds throughput when requests overlap -- and why a sequential single-box
+runtime that benchmarked fine at one user falls over at twenty.
+
+The flip side is operational: vLLM is a service somebody runs (CUDA stack,
+versioned kernels), where the single-box path is one binary. The crossover
+is measured, not guessed: the arrival rate the engagement recorded says
+whether requests overlap, and `fde scan <engagement>` reads it.

{fde_framework-0.1.4 → fde_framework-0.1.6}/framework/templates/retrieval/graph-expanded-retrieval.pgvector.py.j2

@@ -45,7 +45,9 @@ LIMIT %(k)s;
 """
 
 # One hop per query, looped in the application: co-mentioned chunks of the
-# frontier, minus everything already found. Bounded by construction.
+# frontier, minus everything already found. Depth is bounded by the loop and
+# breadth by the pool cap -- depth alone is not a bound, because one hub
+# entity can pull the whole corpus in on a single hop.
 EXPAND_SQL = """
 SELECT DISTINCT m2.chunk_id, c.body
 FROM {table}_mentions m1
@@ -72,6 +74,7 @@ class {{ class_name }}:
         dimensions: int = 768,
         max_hops: int = 2,
         seeds: int = 20,
+        max_pool: int = 200,
     ) -> None:
         self.connection = connection
         self.embed = embed
@@ -80,6 +83,7 @@ class {{ class_name }}:
         self.dimensions = dimensions
         self.max_hops = max_hops
         self.seeds = seeds
+        self.max_pool = max_pool
 
     def schema(self) -> str:
         return SCHEMA_SQL.format(table=self.table, dimensions=self.dimensions)
@@ -102,6 +106,7 @@ class {{ class_name }}:
                     {"frontier": frontier, "found": list(found)},
                 )
                 fresh = cursor.fetchall()
+                fresh = fresh[: max(0, self.max_pool - len(found))]
                 if not fresh:
                     break
                 for row in fresh:

{fde_framework-0.1.4 → fde_framework-0.1.6}/framework/templates/retrieval/graph-expanded-retrieval.plain.py.j2

@@ -20,8 +20,9 @@ approach = "{{ approach }}"
 stack = "{{ stack }}"
 
 TOP_K = 8
-SEEDS = 20     # vector recall opens the graph; it does not need to finish the job
-MAX_HOPS = 2   # each hop multiplies the pool, and two answers most questions
+SEEDS = 20      # vector recall opens the graph; it does not need to finish the job
+MAX_HOPS = 2    # each hop multiplies the pool, and two answers most questions
+MAX_POOL = 200  # depth alone is not a bound: one hub entity can pull the corpus in
 
 # Injected at wiring time: retrieve(query, k) -> seed chunks (each a dict with
 # at least id, text, entities); rerank(query, text) -> relevance, higher is
@@ -54,6 +55,8 @@ class {{ class_name }}:
             ]
             fresh = {c["id"]: {**c, "hop": hop}
                      for c in connected if c["id"] not in pool}
+            room = max(0, MAX_POOL - len(pool))
+            fresh = dict(list(fresh.items())[:room])
             if not fresh:
                 break
             pool.update(fresh)

{fde_framework-0.1.4 → fde_framework-0.1.6}/framework/templates/retrieval/graph-expanded-retrieval.qdrant.py.j2

@@ -24,7 +24,8 @@ stack = "{{ stack }}"
 COLLECTION = "{{ component }}_chunks"
 TOP_K = 8
 SEEDS = 20
-MAX_HOPS = 2  # bounded: each hop multiplies latency, and two answers most questions
+MAX_HOPS = 2   # bounded: each hop multiplies latency, and two answers most questions
+MAX_POOL = 200  # depth alone is not a bound: one hub entity can pull the corpus in
 
 
 class QdrantUnwired(RuntimeError):
@@ -62,6 +63,7 @@ def run(query: str, top_k: int = TOP_K) -> list[dict]:
             for chunk_id in adjacency.get(entity, [])
         ]
         fresh = [c for c in dict.fromkeys(connected) if c not in found]
+        fresh = fresh[: max(0, MAX_POOL - len(found))]
         if not fresh:
             break
         records = client.retrieve(collection_name=COLLECTION, ids=fresh,
@@ -69,7 +71,11 @@ def run(query: str, top_k: int = TOP_K) -> list[dict]:
         for record in records:
             found[str(record.id)] = {"id": str(record.id), "hop": hop,
                                      **(record.payload or {})}
-        frontier = fresh
+        # Only ids the store actually returned: on a corpus that keeps
+        # changing, the adjacency map can name a chunk deleted since the
+        # last rebuild, and walking from a ghost is a KeyError in the exact
+        # scenario this variant exists for.
+        frontier = [str(record.id) for record in records]
     scored = [
         {**c, "rerank_score": rerank(query, c.get("text", ""))}
         for c in found.values()

{fde_framework-0.1.4 → fde_framework-0.1.6}/framework/templates/retrieval/graph-retrieval.qdrant.py.j2

@@ -54,10 +54,10 @@ def run(query: str, top_k: int = TOP_K) -> list[dict]:
     frontier = list(found)
     for hop in range(1, MAX_HOPS + 1):
         connected = [
-            chunk_id
+            neighbour
             for chunk_id in frontier
             for entity in found[chunk_id].get("entities", [])
-            for chunk_id2 in adjacency.get(entity, [])
+            for neighbour in adjacency.get(entity, [])
         ]
         fresh = [c for c in dict.fromkeys(connected) if c not in found]
         if not fresh:

{fde_framework-0.1.4 → fde_framework-0.1.6}/pyproject.toml

@@ -2,7 +2,7 @@
 # Distribution name matches the repository and stays unique on any index;
 # the import package and the CLI command remain the short `fde`.
 name = "fde-framework"
-version = "0.1.4"
+version = "0.1.6"
 description = "A framework for Forward Deployed Engineers: from problem statement to a deployable AI project, every decision traced to a fact."
 readme = "README.md"
 license = {text = "Apache-2.0"}

{fde_framework-0.1.4 → fde_framework-0.1.6}/src/fde/architect.py

@@ -53,7 +53,16 @@ class Architecture:
         return {s: v for s, v in self.licences.items() if copyleft(v)}
 
     def fingerprint(self) -> str:
-        return self.decisions.decided_fingerprint()
+        """Topology and decisions together.
+
+        Two engagements with the same approach set but different topologies
+        build byte-different projects (the boundary code differs), and a
+        fingerprint that called them identical was answering a narrower
+        question than the one it prints beside."""
+        import hashlib
+
+        payload = f"{self.topology}:{self.decisions.decided_fingerprint()}"
+        return hashlib.sha256(payload.encode()).hexdigest()[:16]
 
 
 def architect(

{fde_framework-0.1.4 → fde_framework-0.1.6}/src/fde/cli.py

@@ -592,6 +592,12 @@ def ask(
             continue
 
         dimension_entry = registry.dimensions[question.resolves]
+        # A multi-valued dimension never enters the answer space (a second
+        # value is a peer, not a contradiction), so the space alone would
+        # offer this question again immediately -- asked and answered is
+        # asked and answered for this session; a later session can add more.
+        if dimension_entry.multi_valued or isinstance(answer.value, tuple):
+            passed_on.add(question.resolves)
         answered_values = (
             list(answer.value) if isinstance(answer.value, tuple) else [answer.value]
         )
@@ -1517,9 +1523,15 @@ def build_cmd(
         ]
         report = emit(architecture, out, registry=registry,
                       templates=Path(registry_root) / "templates",
-                      pairs_path=Path(root) / "artifacts" / "pairs.jsonl",
+                      # The engagement's own root, never the raw argument: a
+                      # bare name (`fde build receipts`) resolves through
+                      # _engagement, and building the pairs path from the raw
+                      # argument silently emitted a project with an empty
+                      # golden set while 42 verified pairs sat on disk.
+                      pairs_path=engagement.root / "artifacts" / "pairs.jsonl",
                       waivers=waivers,
-                      overrides=applied_overrides)
+                      overrides=applied_overrides,
+                      baseline=engagement.baseline())
     except BuildRefused as exc:
         typer.echo(f"refused: {exc}", err=True)
         raise typer.Exit(1) from exc
@@ -1554,6 +1566,22 @@ def build_cmd(
         _write_compliance(Path(out), locale)
 
     typer.echo(f"wrote {out}")
+    # The exam's size belongs in the build receipt: an empty golden set
+    # emitted next to sixty pairs on disk once read as a finished build.
+    counts = {}
+    for layer in ("golden", "edge_case", "adversarial"):
+        layer_path = Path(out) / "evals" / f"{layer}.jsonl"
+        counts[layer] = sum(
+            1 for line in layer_path.read_text().splitlines() if line.strip()
+        ) if layer_path.exists() else 0
+    typer.echo(f"evals: {counts['golden']} golden, {counts['edge_case']} edge, "
+               f"{counts['adversarial']} adversarial")
+    if not counts["golden"] and (engagement.root / "artifacts" / "pairs.jsonl").exists():
+        typer.echo(
+            "  pairs exist on the engagement but none reached the golden set -- "
+            "mark checked pairs `verified: true` and re-run `fde samples`",
+            err=True,
+        )
     # The delivery is finishable, and the finishing move is one command.
     holdout_path = engagement.root / "artifacts" / "holdout.jsonl"
     hint = f"fde implement {out}"
@@ -1648,7 +1676,15 @@ def scan_cmd(
     # costing figure, because model releases move monthly.
     from fde.scan import MODEL_GUIDANCE_AS_OF, recommend_local_models
 
-    plan = recommend_local_models(hardware)
+    # The engine choice is a concurrency decision as much as a hardware
+    # one: when the engagement recorded an arrival rate, it flows in.
+    arrival = None
+    if root is not None:
+        try:
+            arrival = _engagement(root).profile.get("arrival_rate")
+        except Exception:  # noqa: BLE001 -- a scan without an engagement still scans
+            arrival = None
+    plan = recommend_local_models(hardware, arrival_per_day=arrival)
     typer.echo(f"\nlocal models (guidance as of {MODEL_GUIDANCE_AS_OF} -- "
                f"releases move monthly, verify before install)")
     typer.echo(f"  runtime: {plan.runtime} -- {plan.runtime_reason}")

{fde_framework-0.1.4 → fde_framework-0.1.6}/src/fde/decide.py

@@ -111,9 +111,22 @@ def decide_component(
             )
             continue
         if not any(holds(c, profile, registry) for c in approach.applies_when):
-            rejected.append(
-                Rejected(approach.id, f"nothing here matches {' or '.join(approach.applies_when)}")
-            )
+            reason = f"nothing here matches {' or '.join(approach.applies_when)}"
+            # An unanswered dimension is a different situation from a
+            # mismatched one: the reader should know one question could
+            # change this rejection, not go looking for a design flaw.
+            open_dimensions = sorted({
+                dimension
+                for condition in approach.applies_when
+                for dimension in _referenced(condition)
+                if profile.get(dimension) is None
+            })
+            if open_dimensions:
+                reason += (
+                    f" -- unanswered: {', '.join(open_dimensions)} "
+                    f"(an answer could admit it)"
+                )
+            rejected.append(Rejected(approach.id, reason))
             # Not ruled out -- just not ruled in. If its applies conditions
             # reference something unanswered, an answer could still admit it.
             still_askable.append(approach)